AI can now generate surprisingly capable code from natural-language instructions, and it seems to work well for prototypes and small projects. However, experienced engineers often say that large applications with many users cannot simply be vibe coded and still require human engineering. I have mostly worked on projects with around 10–15 users, so I have trouble understanding what changes at larger scale. Aside from compliance, what specific technical and practical problems make vibe coding risky for a heavily used application?
4 Answers
You can use AI heavily on a large application, and many teams already do. The important distinction is that it is an implementation assistant, not the person responsible for the system. Humans still need to define the architecture, requirements, testing strategy, operational safeguards, and acceptance criteria. Without that knowledge, you may not notice that the AI solved the wrong problem or introduced a subtle flaw.
As a system grows, the cost of changing it goes up. A generated rewrite that looks cleaner may accidentally remove old behavior that other parts of the system depend on. Large systems contain years of bug fixes, undocumented assumptions, integrations, and data constraints. Vibe coding can still be part of the workflow, but the developer needs enough architectural knowledge to preserve those contracts and review every significant change.
It’s not really the user count by itself. A small script or proof of concept can be judged by whether it works in a few obvious cases. A real product has many interacting requirements: reliability, security, deployment, monitoring, backups, data migrations, performance, incident recovery, and compatibility with existing systems. AI can help implement pieces of that, but someone experienced still has to identify the pieces and make sure they fit together.
AI is very good at producing code that passes a quick test, but “it works” does not necessarily mean it is safe or correct. It might miss authorization checks, mishandle concurrent requests, leak data, or fail badly under unusual input. An experienced engineer can review the generated code, recognize the likely failure modes, and design tests for cases the original prompt never mentioned.
That’s why the strongest argument is not “AI can never build large systems.” It’s that a person who cannot understand, test, and review the generated work cannot safely be the final owner of one.